Research map / Video / Video-language understanding
Text-Conditioned Resampler For Long Form Video Understanding
Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, Federico Tombari
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add task-conditioning text to the resampler (ablation) | TCR · NextQA temporal questions | accuracy: 61.1 -> 64.9 | Improved |
| corrupt the temporal word in conditioning (before/after swap) (ablation) | TCR · NextQA temporal questions | accuracy: 64.9 -> 55.3 | Hurt |
| add initialisation stage (contrastive and matching) (ablation) | TCR · NextQA val (short schedule) | accuracy: 63.3 -> 66.1 | Improved |
| add pre-training with LLM on YTT-1B (ablation) | TCR · NextQA val (short schedule) | accuracy: 52.1 -> 66.1 | Improved |
| add denoising pre-training task (iii) (ablation) | TCR pre-training · NextQA val (short schedule) | accuracy: 65.6 -> 66.1 | Improved |
| increase input frames from 32 to 92 (ablation) | TCR · NextQA val (short schedule) | accuracy: 64.4 -> 66.2 | Improved |
| increase input frames from 92 to 124 (ablation) | TCR · NextQA val (short schedule) | accuracy: 66.2 -> 65.9 | No clear effect |
| use 128 queries instead of 32 or 256 (ablation) | TCR · NextQA val (short schedule) | accuracy: 62.7 / 64.3 -> 66.2 | Mixed |
| add task-specific special tokens (ablation) | TCR · NextQA val | accuracy: 72.7 -> 73.5 | Improved |
| increase observed frames from 8 to 92 with TCR (vs BLIP2 averaging) | BLIP2 · EgoSchema | QA accuracy (%): 27.2 -> 34.2 | Improved |
| observe the whole video instead of only before the last action (ablation) | TCR · Ego4D LTA val | ActionED: 0.8792 -> 0.8522 | Improved |
| TCR generative LLM approach vs ASL specialist on moment queries | TCR · Ego4D Moment Queries val | avg mAP: 27.85 -> 25.45 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- PALM* — PALM: Predicting Actions through Language Models (ECCV 2024) · better on 1 of 1 shared result
- SCSampler — SCSampler: Sampling Salient Clips From Video for Efficient Action Recognition (ICCV 2019) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations (AAAI 2025 · similarity 0.86)
- Mimir: Improving Video Diffusion Models for Precise Text Understanding (CVPR 2025 · similarity 0.86)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.85)
- Frozen Transformers in Language Models Are Effective Visual Encoder Layers (ICLR 2024 · similarity 0.85)
- A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames (CVPR 2024 · similarity 0.85)
- AdaCM^2: On Understanding Extremely Long-Term Video with Adaptive Cross-Modality Memory Reduction (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
